Four follow-ups to the previous commit, three of them cases where the new
runner still did the thing it was written to stop doing.
A build that was already up to date ran nothing. `make` says "up to date" on
the second invocation, no executable changes mtime, the before/after snapshot
diff is empty -- and the runner printed a line and stopped, which is exactly
the complaint that started this. It now falls back to running the executables
that are there. Proved with a scratch course whose makefile has a real file
target, because both make courses in this repo rebuild unconditionally and
neither can reach the branch.
One broken target aborted the rest. `run <dir> all` stopped at the first
failure under set -e, so ELAC's second document was never attempted. Failures
are collected per target and reported together at the end instead.
An index was ignored when a course had one target: `run ECOAR 5` silently ran
target 1. Selections are validated before the single-target shortcut now.
And "ECOAR has 1 targets" read like a bug in the tool.
That last fix then falsified a claim in TODO-refactor-remaining.md. It said
ELAC's projectA and AIS's report/final/ver1 "fail with latexmk exit 12" and
were broken documents. With build/ deleted first, both courses build every
document and exit 0. The exit 12 reproduced once against pre-existing build
state and never from clean; the note now says that, and says which part is
inference rather than re-verified, because the original failing state is gone.
All 44 probes still produce a byte-identical runnability-local.tsv.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172dj34FDpao6wsix8ZuCzE
Reported by kuhy: "I run run.sh and I have no idea where the actual program is
and how to run it". In EPFU that was one `gcc -Wall -Wextra -pedantic main.c`
line and nothing else -- the makefile has no -o, so the binary is an a.out that
nothing mentions, and it was never executed.
Two separate failures, both fixed.
Building is not running. make, cxx, maven and dotnet built and stopped; make,
maven and dotnet did not even print where the output went, while cxx and tex
did. They all execute what they produced now. `make` is the awkward one,
because a makefile need not say what it writes: the runner snapshots the
executables either side of the build and runs whatever is new.
A course is not always one program. Twelve of them are not -- EPFU a game and
a lab, EOOP six binaries, EARIN eight labs, MOM three reports. With no target
named, an interactive run prints a numbered menu and asks; `run <dir> 2`, a
path substring, or `all` skips the prompt. Enumeration deliberately scans the
course root rather than --entry, because EPFU pinning --entry to penguins/src
is precisely why labs/krudnic3_lab1.c had never been compiled. --entry is the
default now, not the only.
Interactive by default: EPFU's penguins and ECOAR's puzzle both read stdin,
and you typed `run`, so you are at the keyboard. --batch (and any
non-terminal stdin) means no prompt, stdin from /dev/null and a 60s cap per
program, so an unattended run cannot wedge. --probe never reaches any of this
and does not even enumerate targets: all 44 probes produce a byte-identical
gates/runnability-local.tsv, and `status` is still seconds long.
Running things immediately found two bugs that could not have surfaced before,
both recorded in TODO-refactor-remaining.md rather than fixed here, since both
mean touching a submitted .c:
- EPFU/labs/krudnic3_lab1.c dies with SIGFPE, integer division by zero.
- ECOAR's puzzle.out writes dest.bmp beside its source, so the tree guard
fails the run. That guard is doing exactly its job.
Verified: EPFU both targets, EOOP under a pty answering the prompt, EADS,
ECOAR, MOM. Full gate green, shellcheck clean on all 44 runners.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172dj34FDpao6wsix8ZuCzE
